SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.

A Thorough Analysis of How to Create 'Non-AI-Looking Images'✨ Introducing how to create subtle AI images using PixAI—a must-read for AI manga artists and AI illustrators!

ハカセアイ (Ai-Hakase)🐱YouTube × 記事でトレンドAIを解説する為のnote🐾

2025年7月4日 20:38

Hello everyone! I'm the type of person who is obsessed with the millimeter-perfect placement of my pillow before sleep, Ai Hakase!

First of all, the content of this Note article can also be viewed in the YouTube video version, so please check it out here!

It might be sudden, but in this day and age, with the power of AI, you can get beautiful, professional-looking illustrations in an instant with just the push of a button...

However, don't you feel like lately, it's becoming easier to tell at a glance, 'Oh, this is AI art'?

I'm sure many of you are thinking:

'I want to create more original art that feels like my own!' 'I want to get rid of that AI look and try my hand at creating soft, emotional, "subtle art"!'

I suspect many of you feel that way.

This time, I would like to introduce how to create images that don't look like AI, focusing mainly on illustrations and incorporating some quite professional perspectives!

An easy method using PixAI

Actually, trying to do this with ComfyUI or SDWebUI can be quite difficult and tedious, but it is known that you can create high-quality images quite easily and affordably using an image generation AI service that can be used on the web from smartphones or PCs called "PixAI", so I will explain how to use that.

I will thoroughly explain how to create illustrations that don't look like AI scientifically, while touching on recent innovations in image generation AI technology, such as the concepts of models, sampler algorithms, and LoRA. If you know this, you'll be able to create things that make people ask, 'Did an AI really draw this?' much faster than others!

In fact, I've been receiving quite a few consulting requests for this from both individuals and companies lately, and I think this content will be very useful for many people, especially AI manga artists and AI illustrators, so please watch until the end!

What we will cover today

So, here is what we will be talking about today!

【① The key to removing the AI look! Scientific selection of "models" and "samplers"】
【② The magic of creating your own art style! The golden ratio of "LoRA" roles and weight adjustment】
【③ Infinite generation for almost free!? Revealing the secret tricks to save credits in PixAI!】

Also, since I am converting the content of this note into videos using AI and publishing them on YouTube, if you haven't already, please follow this note and subscribe to the YouTube channel as well!

https://note.com/ai_hakase

Furthermore, the public version will be accessible from this link as soon as the video rendering is complete, so please take a look if you'd like!

https://www.youtube.com/%40AI-Hakase

With that, thank you for your support today as well!

How to choose models, samplers, and LoRAs

So, first, I'd like to introduce the most important points, such as how to choose models, samplers, and LoRAs, and share with you some quite efficient and very useful testing methods unique to PixAI!

First, as an overall flow, the basic process is deciding on a model → deciding on a sampler → deciding on a LoRA → and then adding flavor using prompts and additional LoRAs.

Once you get used to it, it's convenient to have a few templates of your own, but this flow is very important!

I will explain this in detail later, but some people make the mistake of immediately looking for a LoRA that gives a certain style first, only to find they can't create the image they want at all. The basic idea is that the model is the main dish that determines the basic art style, and since a LoRA is just additional training that adds the quality of the images learned by the LoRA onto the base model's image generation, it's better to steadily start from this model selection stage!

So, I will explain them one by one, so please remember them well!

How to choose a model

Differences between illustration-style and Noob-style

So, first, let's talk about models.

There are really this many models available on PixAI now, including those that are independently trained or open-source, but they can be broadly divided into illustration-style and Noob-style.

Many of you might be wondering what those are, but simply put, illustration-style is an image with a shiny feel and clear lines, and Noob-style has the atmosphere of something hand-drawn by a human, which is subtle and properly trained on the work of actual illustrators.

Therefore, basically, if you want to create something that doesn't look like AI, using Noob-style is the key!

In this example, Haruka V2 is illustration-style, and Animagine XL is Noob-style.

How to distinguish models

So, the question is how to actually distinguish them, but this is a bit complicated; in fact, there are quite a few models out there now that are a merge of the illustration-style and Noob-style I just showed you.

So, frankly speaking, you don't need to distinguish them!

First, just choose a few that you like from models with subtle illustrations like this!

Selecting a sampling method

What is a sampling method?

Once the model is decided, we will move on to selecting the sampling method.

As for what a sampling method is, as shown in the list of sampling methods under Advanced Settings below the model, it is what determines the various algorithms the AI uses when generating images!

You could run through these one by one, but there is a more efficient way, and basically, you only need to select and try these 4 items circled in red here!

Comparison of Sampling Methods

I made a demo with HARUKA V2 for now, and as you can see, the art style generated changes completely depending on which one you choose!

And this is the pattern for Otome V2. I quite like Euler or Euler A, so I use them often, but please choose your favorite and decide on a sampling algorithm first!

Efficient Testing Method

Regarding the experimental method for this, there is a secret trick to try generating these images for just 200 credits, so please use that to test efficiently.

As for how to do it, basically you use this DMD2 LoRA to achieve high-speed generation and credit savings.

For details, I explain this in the Tips section at the end of my YouTube video, so please take a look at that.

Basically, select the LoRA and

set the steps between 8 and 12, and adjust the CFG scale to 1.1.

With this, you can experiment with sampling algorithms for just 200 credits, so it's almost like testing for free!

Points to Note with Noob-based Models

By the way, if the training content of the model you select is based on Noob models, you may see artifacts like this.

In that case, please apply this LoRA instead and change the sampler to Euler A!

This will allow for high-speed generation specialized for Noob models, and you can generate images without artifacts as shown here.

However, in that case, it seems other sampling algorithms will cause artifacts again, so if you want to use a sampling algorithm other than Euler A, the process would be to use about 2000 credits as usual to generate!

By the way, regarding the selection of this sampling method, if you try it and decide you want to test a different model, try the other model using the same procedure, and first make sure to solidify the set of sampling method and model!

Deciding the Style with LoRA

Once you reach this point, it finally feels like it is time to decide on the style with LoRA.

As you have seen so far, the selection of the model and the sampling algorithm stage already make it look quite convincing.

I will add LoRA here to generate it, and enhance it to make the picture look more natural like this!

How to find LoRA

First, regarding how to find LoRA, basically, please select LoRA and SDXL using the filter function on this model screen!

https://pixai.art/ja/market?type=ANY_SDXL_LORA&refCode=MB35FD9N

Then, a list of usable LoRA should appear on the screen like this.

How to use LoRA and tips

Basic concept of LoRA

Also, as I mentioned at the beginning, LoRA is something like a 'seasoning' that gives additional instructions such as 'specific art style,' 'specific character,' or 'specific color tone' to the base model (Haruka v2, Otome V2, etc.), but there are 'theories of thinking' and 'adjustment tips' to get closer to the desired result, so I would like to introduce them.

Types of LoRA

First, regarding the types of LoRA, they can be broadly divided into the following, but it is very easy to understand if you think of them as a 'paint set'.

Art Style LoRA : Paints that change the touch of the entire picture, such as 'watercolor style' or 'sketch style'.
Color LoRA : Paints that change the atmosphere of the color usage, such as 'pastel colors' or 'retro color tones'.
Quality LoRA : Something like a transparent varnish that enhances the perfection of the picture, such as 'detail up' or 'noise removal'.
Character LoRA : Special paints for reproducing specific characters.

Recommended LoRA

For now, my recommendations are around here, but I think it would be good for you to also keep an eye on a few of your favorites!

For details, please try searching with keywords like this on the search screen mentioned earlier!

Keywords to make line art delicate : `line art`, `sketch`, `delicate`, `thin lines`
Keywords to make coloring pale : `watercolor`, `pastel color`, `light color`
Search by style name : Searching by the names of famous illustrators or specific art styles (e.g., `Qiqu Style`) is also effective!

The Golden Ratio for Weight Adjustment

Once you've decided what you want to do, you'll apply it immediately, but there's a concept called weight , so I'll explain that a bit!

What is this weight, you ask? It is the value that determines how strongly the LoRA style's flavor is applied!

In other words, the higher you set it, the more likely those characteristics are to appear!

Basically, the recommended settings are written on the model page, so you can just follow those, but I'll introduce a few golden ratios .

※ However, there is no absolute correct logic that says 'this is the only way' , so I would appreciate it if you could think about it flexibly.

Basically, please think of it like this!

① Main Art Style LoRA (Weight: 0.6–0.9)
This is the protagonist that determines the skeleton of the picture! It is the most important LoRA that embodies your will, saying, 'I want this picture to have this touch!' Set it a little higher. The basic approach is to try it at around `0.8` at first, then lower it if it's too strong, or raise it if it's too weak!

② Support LoRA (Weight: 0.3–0.6)
The great supporting actor that helps the main art style! Its role is to 'gently add' color and atmosphere. The trick is to keep the weight modest so as not to ruin the main flavor...! For example, when you want to 'keep the delicate line art of Qiqu Style as is, but make the colors cute pastel colors,' you would set it to around 0.4!

③ Auxiliary LoRA (Weight: 0.8–1.0)
The 'unsung hero' that raises the overall quality. Things like detail enhancement that don't easily clash with the art style can often be set higher without issue.

Important Point

As a point, if you set the weights of multiple LoRAs all high (e.g., all 0.8 or higher), their respective characteristics will clash, and the picture will easily fall apart (the so-called 'yakisoba phenomenon').

So, although it's just a guideline, the trick is to decide on one main one and keep the others modest . It's good to have a sense that when the total weight value exceeds `2.0` to `2.5`, the risk of collapse increases!

By the way, if you set the ones you aren't using to 0 like this, that part won't be applied at all, which is convenient when testing!

Tips for Testing

And when testing LoRAs, basically, I think it's good to fix something called a seed value so that changes are easier to understand!

The trick is a cycle like fixing the seed value → trying a few main art style LoRAs → finding a good one and then trying support LoRAs → and then trying auxiliary LoRAs after that!

Prompt Settings

Next are the positive prompt and negative prompt sections.

Basically, it's good to include words in the prompt that suggest thin lines, such as "(thin lines:1.2)".

For the negative prompt, it's good to include things that exclude thick lines or overly sharp looks, such as
"(thick lines:1.3), heavy shadowing, oversaturated, 3d"
.

Actual Production Examples

So, as a summary of everything up to this point, let's take a look at how things change as we add LoRAs one by one!

First, we talked about deciding on the model and sampler!

For this, we will use the Peluche MIX model.

https://pixai.art/ja/model/1830832313621573017?refCode=MB35FD9N

As for the sampler, I'd like to use Euler A!

Next is the LoRA configuration. Since I am currently subscribed to the PixAI Plus plan,

https://pixai.art/ja/membership/plans?refCode=MB35FD9N

I'd like to test these LoRAs one by one, using up to the limit of 10, and setting the weight to 0 for the ones I'm not using!

By the way, with these settings, it costs 200 credits to use, so even if you generate 10 times using each LoRA once, it only costs 2000 credits!

So, first is the state without any LoRA. This is the model's original flavor. It's cute in its own way, but it still feels a bit "AI-like"...

From here, I will first add only the main art style LoRA with a weight of 0.8!

It's gotten much better, but next I'll add things like color using a support LoRA!

And finally, I'll sharpen the fine details using a quality-focused LoRA!

Also, if you use something like Hires. fix to upscale it, I think you'll get an even more beautiful image!

It looks quite different compared to the beginning, doesn't it!

PixAI Pricing Plans

As you can see, PixAI itself is quite usable for free, but as you've noticed, it is basically credit-based, so the membership plan starting at $9.99 per month is highly recommended as it offers great value!

https://pixai.art/membership/plans?refCode=MB35FD9N

As you can see, depending on the plan, a large number of credits are granted at the beginning of the month, but even with the cheapest starter plan, combined with the daily credits, it calculates to about 600,000 to 700,000 credits accumulated per month!

Therefore, when combined with the credit-saving features I mentioned earlier, it feels like you can generate images almost infinitely!

You can check the detailed benefits for the latest version via this link, so please take a look if you'd like!

https://pixai.art/membership/plans?refCode=MB35FD9N

Summary

How was it! To summarize this time, the quality of AI illustrations can be dramatically improved by anyone depending on the combination of the base model and the magic spice called LoRA!

1. The first step to erasing the 'AI look' is choosing the right 'model' and 'sampler'!

First, it is super important to choose a model that serves as the foundation for your ideal art style, and try out several samplers, which are drawing algorithms, to find the combination you like best!

2. For LoRA, 'division of roles' and 'weight adjustment' are everything!

By assigning roles to LoRA such as 'main art style,' 'supporting colors,' and 'auxiliary quality,' and adjusting the weights, you can create the delicate art style you are aiming for!

3. Test efficiently with PixAI's 'credit-saving techniques'!

If you use tricks like high-speed generation LoRA, you can try many patterns almost for free. Keep experimenting and find your own golden ratio!

Please use these techniques to create many of your own original works!

Did you enjoy this article and video?

So, what did you think of this article and video? If you found it helpful, I would appreciate it if you could follow, like, and hit the thumbs-up button on Note as well.